MEC-Cox: ML boosts survival trial analysis with calibrated weights

A new paper on arXiv (2606.08305) introduces MEC-Cox, a method that combines machine learning with generalized entropy calibration to improve hazard-ratio estimation in externally controlled survival trials. Developed by Se Yoon Lee, Yonghyun Kwon, and Jae Kwang Kim, MEC-Cox targets the average treatment effect on the treated (ATT) marginal hazard ratio—comparing a treated trial population to a counterfactual control arm when concurrent randomized controls are unavailable. This scenario is common in oncology and rare diseases. The approach begins with normalized source-propensity-score odds weights for external controls, then applies Bregman calibration to balance cross-fitted prognostic summaries (e.g., survival predictions from Cox models, penalized models, or ML) between external controls and treated trial patients. The resulting weights serve dual roles: transporting the external control distribution and balancing prognostic scores.

MEC-Cox builds on the machine-learning-assisted generalized entropy calibration (MEC) framework by the same authors (2026), extending it to IPW Cox regression—a setting where weighting affects both event contributions and risk-set averages, complicating direct ML integration. The paper establishes consistency, characterizes an efficiency gain from calibration, and provides a stacked sandwich variance estimator for valid inference. Simulations demonstrate that MEC-Cox reduces bias, increases efficiency, and improves coverage compared to standard IPW Cox regression. This method offers a principled, flexible way to incorporate flexible machine learning into survival analysis for causal inference, potentially improving decision-making in drug development and regulatory evaluations where external controls are used.